nsz Git - libfirm/blob - ir/be/becopyopt.c

   1 /*
   2  * This file is part of libFirm.
   3  * Copyright (C) 2012 University of Karlsruhe.
   4  */
   5
   6 /**
   7  * @file
   8  * @brief       Copy minimization driver.
   9  * @author      Daniel Grund
  10  * @date        12.04.2005
  11  *
  12  * Main file for the optimization reducing the copies needed for:
  13  * - Phi coalescing
  14  * - Register-constrained nodes
  15  * - Two-address code instructions
  16  */
  17 #include "config.h"
  18
  19 #include "debug.h"
  20 #include "error.h"
  21 #include "execfreq_t.h"
  22 #include "irdump_t.h"
  23 #include "iredges_t.h"
  24 #include "irgraph.h"
  25 #include "irgwalk.h"
  26 #include "irloop_t.h"
  27 #include "irnode.h"
  28 #include "irprintf.h"
  29 #include "irprog.h"
  30 #include "irtools.h"
  31 #include "pmap.h"
  32 #include "raw_bitset.h"
  33 #include "util.h"
  34 #include "xmalloc.h"
  35
  36 #include "bearch.h"
  37 #include "becopyopt_t.h"
  38 #include "becopystat.h"
  39 #include "bedump.h"
  40 #include "beifg.h"
  41 #include "beinsn_t.h"
  42 #include "beintlive_t.h"
  43 #include "beirg.h"
  44 #include "belive_t.h"
  45 #include "bemodule.h"
  46 #include "benode.h"
  47 #include "besched.h"
  48 #include "statev_t.h"
  49 #include "beutil.h"
  50
  51 #include "lc_opts.h"
  52 #include "lc_opts_enum.h"
  53
  54 #define DUMP_BEFORE 1
  55 #define DUMP_AFTER  2
  56 #define DUMP_APPEL  4
  57 #define DUMP_ALL    2 * DUMP_APPEL - 1
  58
  59 #define COST_FUNC_FREQ     1
  60 #define COST_FUNC_LOOP     2
  61 #define COST_FUNC_ALL_ONE  3
  62
  63 /**
  64  * Flags for dumping the IFG.
  65  */
  66 enum {
  67         CO_IFG_DUMP_COLORS = 1 << 0, /**< Dump the graph colored. */
  68         CO_IFG_DUMP_LABELS = 1 << 1, /**< Dump node/edge labels. */
  69         CO_IFG_DUMP_SHAPE  = 1 << 2, /**< Give constrained nodes special shapes. */
  70         CO_IFG_DUMP_CONSTR = 1 << 3, /**< Dump the node constraints in the label. */
  71 };
  72
  73 static int co_get_costs_loop_depth(const ir_node *root, int pos);
  74 static int co_get_costs_exec_freq(const ir_node *root, int pos);
  75 static int co_get_costs_all_one(const ir_node *root, int pos);
  76
  77 static unsigned   dump_flags  = 0;
  78 static unsigned   style_flags = CO_IFG_DUMP_COLORS;
  79 static int        do_stats    = 0;
  80 static cost_fct_t cost_func   = co_get_costs_exec_freq;
  81 static int        improve     = 1;
  82
  83 static const lc_opt_enum_mask_items_t dump_items[] = {
  84         { "before",  DUMP_BEFORE },
  85         { "after",   DUMP_AFTER  },
  86         { "appel",   DUMP_APPEL  },
  87         { "all",     DUMP_ALL    },
  88         { NULL,      0 }
  89 };
  90
  91 static const lc_opt_enum_mask_items_t style_items[] = {
  92         { "color",   CO_IFG_DUMP_COLORS },
  93         { "labels",  CO_IFG_DUMP_LABELS },
  94         { "constr",  CO_IFG_DUMP_CONSTR },
  95         { "shape",   CO_IFG_DUMP_SHAPE  },
  96         { "full",    2 * CO_IFG_DUMP_SHAPE - 1 },
  97         { NULL,      0 }
  98 };
  99
 100 typedef int (*opt_funcptr)(void);
 101 static const lc_opt_enum_func_ptr_items_t cost_func_items[] = {
 102         { "freq",   (opt_funcptr) co_get_costs_exec_freq },
 103         { "loop",   (opt_funcptr) co_get_costs_loop_depth },
 104         { "one",    (opt_funcptr) co_get_costs_all_one },
 105         { NULL,     NULL }
 106 };
 107
 108 static lc_opt_enum_mask_var_t dump_var = {
 109         &dump_flags, dump_items
 110 };
 111
 112 static lc_opt_enum_mask_var_t style_var = {
 113         &style_flags, style_items
 114 };
 115
 116 static lc_opt_enum_func_ptr_var_t cost_func_var = {
 117         (opt_funcptr*) &cost_func, cost_func_items
 118 };
 119
 120 static const lc_opt_table_entry_t options[] = {
 121         LC_OPT_ENT_ENUM_FUNC_PTR ("cost",    "select a cost function",                                  &cost_func_var),
 122         LC_OPT_ENT_ENUM_MASK     ("dump",    "dump ifg before or after copy optimization",              &dump_var),
 123         LC_OPT_ENT_ENUM_MASK     ("style",   "dump style for ifg dumping",                              &style_var),
 124         LC_OPT_ENT_BOOL          ("stats",   "dump statistics after each optimization",                 &do_stats),
 125         LC_OPT_ENT_BOOL          ("improve", "run heur1 before if algo can exploit start solutions",    &improve),
 126         LC_OPT_LAST
 127 };
 128
 129 static be_module_list_entry_t *copyopts = NULL;
 130 static const co_algo_info *selected_copyopt = NULL;
 131
 132 void be_register_copyopt(const char *name, co_algo_info *copyopt)
 133 {
 134         if (selected_copyopt == NULL)
 135                 selected_copyopt = copyopt;
 136         be_add_module_to_list(&copyopts, name, copyopt);
 137 }
 138
 139 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copyopt)
 140 void be_init_copyopt(void)
 141 {
 142         lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
 143         lc_opt_entry_t *ra_grp = lc_opt_get_grp(be_grp, "ra");
 144         lc_opt_entry_t *chordal_grp = lc_opt_get_grp(ra_grp, "chordal");
 145         lc_opt_entry_t *co_grp = lc_opt_get_grp(chordal_grp, "co");
 146
 147         lc_opt_add_table(co_grp, options);
 148         be_add_module_list_opt(co_grp, "algo", "select copy optimization algo",
 149                                        &copyopts, (void**) &selected_copyopt);
 150 }
 151
 152 static int void_algo(copy_opt_t *co)
 153 {
 154         (void) co;
 155         return 0;
 156 }
 157
 158 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copynone)
 159 void be_init_copynone(void)
 160 {
 161         static co_algo_info copyheur = {
 162                 void_algo, 0
 163         };
 164
 165         be_register_copyopt("none", &copyheur);
 166 }
 167
 168 #undef QUICK_AND_DIRTY_HACK
 169
 170 static int nodes_interfere(const be_chordal_env_t *env, const ir_node *a, const ir_node *b)
 171 {
 172         if (env->ifg)
 173                 return be_ifg_connected(env->ifg, a, b);
 174         else {
 175                 be_lv_t *lv = be_get_irg_liveness(env->irg);
 176                 return be_values_interfere(lv, a, b);
 177         }
 178 }
 179
 180
 181 /******************************************************************************
 182     _____                           _
 183    / ____|                         | |
 184   | |  __  ___ _ __   ___ _ __ __ _| |
 185   | | |_ |/ _ \ '_ \ / _ \ '__/ _` | |
 186   | |__| |  __/ | | |  __/ | | (_| | |
 187    \_____|\___|_| |_|\___|_|  \__,_|_|
 188
 189  ******************************************************************************/
 190
 191 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
 192
 193
 194 copy_opt_t *new_copy_opt(be_chordal_env_t *chordal_env, cost_fct_t get_costs)
 195 {
 196         const char *s1, *s2, *s3;
 197         size_t len;
 198         copy_opt_t *co;
 199
 200         FIRM_DBG_REGISTER(dbg, "ir.be.copyopt");
 201
 202         co = XMALLOCZ(copy_opt_t);
 203         co->cenv      = chordal_env;
 204         co->irg       = chordal_env->irg;
 205         co->cls       = chordal_env->cls;
 206         co->get_costs = get_costs;
 207
 208         s1 = get_irp_name();
 209         s2 = get_entity_name(get_irg_entity(co->irg));
 210         s3 = chordal_env->cls->name;
 211         len = strlen(s1) + strlen(s2) + strlen(s3) + 5;
 212         co->name = XMALLOCN(char, len);
 213         snprintf(co->name, len, "%s__%s__%s", s1, s2, s3);
 214
 215         return co;
 216 }
 217
 218 void free_copy_opt(copy_opt_t *co)
 219 {
 220         xfree(co->name);
 221         free(co);
 222 }
 223
 224 /**
 225  * Checks if a node is optimizable, viz. has something to do with coalescing
 226  * @param irn  The irn to check
 227  */
 228 static int co_is_optimizable_root(ir_node *irn)
 229 {
 230         arch_register_req_t const *const req = arch_get_irn_register_req(irn);
 231         if (arch_register_req_is(req, ignore))
 232                 return 0;
 233
 234         if (is_Reg_Phi(irn) || is_Perm_Proj(irn))
 235                 return 1;
 236
 237         if (arch_register_req_is(req, should_be_same))
 238                 return 1;
 239
 240         return 0;
 241 }
 242
 243 /**
 244  * Computes the costs of a copy according to loop depth
 245  * @param pos  the argument position of arg in the root arguments
 246  * @return     Must be >= 0 in all cases.
 247  */
 248 static int co_get_costs_loop_depth(const ir_node *root, int pos)
 249 {
 250         ir_node *block = get_nodes_block(root);
 251         ir_loop *loop;
 252         int      cost;
 253
 254         if (is_Phi(root)) {
 255                 block = get_Block_cfgpred_block(block, pos);
 256         }
 257         loop = get_irn_loop(block);
 258         if (loop) {
 259                 int d = get_loop_depth(loop);
 260                 cost = d*d;
 261         } else {
 262                 cost = 0;
 263         }
 264         return 1+cost;
 265 }
 266
 267 static ir_execfreq_int_factors factors;
 268
 269 /**
 270  * Computes the costs of a copy according to execution frequency
 271  * @param pos  the argument position of arg in the root arguments
 272  * @return Must be >= 0 in all cases.
 273  */
 274 static int co_get_costs_exec_freq(const ir_node *root, int pos)
 275 {
 276         ir_node *root_bl = get_nodes_block(root);
 277         ir_node *copy_bl
 278                 = is_Phi(root) ? get_Block_cfgpred_block(root_bl, pos) : root_bl;
 279         int      res     = get_block_execfreq_int(&factors, copy_bl);
 280
 281         /* don't allow values smaller than one. */
 282         return res < 1 ? 1 : res;
 283 }
 284
 285 /**
 286  * All costs equal 1. Using this will reduce the _number_ of copies.
 287  * @param co   The copy opt object.
 288  * @return Must be >= 0 in all cases.
 289  */
 290 static int co_get_costs_all_one(const ir_node *root, int pos)
 291 {
 292         (void) root;
 293         (void) pos;
 294         return 1;
 295 }
 296
 297 /******************************************************************************
 298    ____        _   _    _       _ _          _____ _
 299   / __ \      | | | |  | |     (_) |        / ____| |
 300  | |  | |_ __ | |_| |  | |_ __  _| |_ ___  | (___ | |_ ___  _ __ __ _  __ _  ___
 301  | |  | | '_ \| __| |  | | '_ \| | __/ __|  \___ \| __/ _ \| '__/ _` |/ _` |/ _ \
 302  | |__| | |_) | |_| |__| | | | | | |_\__ \  ____) | || (_) | | | (_| | (_| |  __/
 303   \____/| .__/ \__|\____/|_| |_|_|\__|___/ |_____/ \__\___/|_|  \__,_|\__, |\___|
 304         | |                                                            __/ |
 305         |_|                                                           |___/
 306  ******************************************************************************/
 307
 308 /**
 309  * Determines a maximum weighted independent set with respect to
 310  * the interference and conflict edges of all nodes in a qnode.
 311  */
 312 static int ou_max_ind_set_costs(unit_t *ou)
 313 {
 314         be_chordal_env_t *chordal_env = ou->co->cenv;
 315         ir_node **safe, **unsafe;
 316         int i, o, safe_count, safe_costs, unsafe_count, *unsafe_costs;
 317         bitset_t *curr;
 318         int curr_weight, best_weight = 0;
 319
 320         /* assign the nodes into two groups.
 321          * safe: node has no interference, hence it is in every max stable set.
 322          * unsafe: node has an interference
 323          */
 324         safe         = ALLOCAN(ir_node*, ou->node_count - 1);
 325         safe_costs   = 0;
 326         safe_count   = 0;
 327         unsafe       = ALLOCAN(ir_node*, ou->node_count - 1);
 328         unsafe_costs = ALLOCAN(int,      ou->node_count - 1);
 329         unsafe_count = 0;
 330         for (i=1; i<ou->node_count; ++i) {
 331                 int is_safe = 1;
 332                 for (o=1; o<ou->node_count; ++o) {
 333                         if (i==o)
 334                                 continue;
 335                         if (nodes_interfere(chordal_env, ou->nodes[i], ou->nodes[o])) {
 336                                 unsafe_costs[unsafe_count] = ou->costs[i];
 337                                 unsafe[unsafe_count] = ou->nodes[i];
 338                                 ++unsafe_count;
 339                                 is_safe = 0;
 340                                 break;
 341                         }
 342                 }
 343                 if (is_safe) {
 344                         safe_costs += ou->costs[i];
 345                         safe[safe_count++] = ou->nodes[i];
 346                 }
 347         }
 348
 349
 350         /* now compute the best set out of the unsafe nodes*/
 351         if (unsafe_count > MIS_HEUR_TRIGGER) {
 352                 bitset_t *best = bitset_alloca(unsafe_count);
 353                 /* Heuristic: Greedy trial and error form index 0 to unsafe_count-1 */
 354                 for (i=0; i<unsafe_count; ++i) {
 355                         bitset_set(best, i);
 356                         /* check if it is a stable set */
 357                         for (o=bitset_next_set(best, 0); o!=-1 && o<i; o=bitset_next_set(best, o+1))
 358                                 if (nodes_interfere(chordal_env, unsafe[i], unsafe[o])) {
 359                                         bitset_clear(best, i); /* clear the bit and try next one */
 360                                         break;
 361                                 }
 362                 }
 363                 /* compute the weight */
 364                 bitset_foreach(best, pos)
 365                         best_weight += unsafe_costs[pos];
 366         } else {
 367                 /* Exact Algorithm: Brute force */
 368                 curr = bitset_alloca(unsafe_count);
 369                 bitset_set_all(curr);
 370                 while (bitset_popcount(curr) != 0) {
 371                         /* check if curr is a stable set */
 372                         for (i=bitset_next_set(curr, 0); i!=-1; i=bitset_next_set(curr, i+1))
 373                                 for (o=bitset_next_set(curr, i+1); o!=-1; o=bitset_next_set(curr, o+1)) /* !!!!! difference to qnode_max_ind_set(): NOT (curr, i) */
 374                                                 if (nodes_interfere(chordal_env, unsafe[i], unsafe[o]))
 375                                                         goto no_stable_set;
 376
 377                         /* if we arrive here, we have a stable set */
 378                         /* compute the weight of the stable set*/
 379                         curr_weight = 0;
 380                         bitset_foreach(curr, pos)
 381                                 curr_weight += unsafe_costs[pos];
 382
 383                         /* any better ? */
 384                         if (curr_weight > best_weight) {
 385                                 best_weight = curr_weight;
 386                         }
 387
 388         no_stable_set:
 389                         bitset_minus1(curr);
 390                 }
 391         }
 392
 393         return safe_costs+best_weight;
 394 }
 395
 396 static void co_collect_units(ir_node *irn, void *env)
 397 {
 398         const arch_register_req_t *req;
 399         copy_opt_t                *co  = (copy_opt_t*)env;
 400         unit_t *unit;
 401
 402         if (get_irn_mode(irn) == mode_T)
 403                 return;
 404         req = arch_get_irn_register_req(irn);
 405         if (req->cls != co->cls)
 406                 return;
 407         if (!co_is_optimizable_root(irn))
 408                 return;
 409
 410         /* Init a new unit */
 411         unit = XMALLOCZ(unit_t);
 412         unit->co = co;
 413         unit->node_count = 1;
 414         INIT_LIST_HEAD(&unit->queue);
 415
 416         /* Phi with some/all of its arguments */
 417         if (is_Reg_Phi(irn)) {
 418                 int i, arity;
 419
 420                 /* init */
 421                 arity = get_irn_arity(irn);
 422                 unit->nodes = XMALLOCN(ir_node*, arity + 1);
 423                 unit->costs = XMALLOCN(int,      arity + 1);
 424                 unit->nodes[0] = irn;
 425
 426                 /* fill */
 427                 for (i=0; i<arity; ++i) {
 428                         int o, arg_pos;
 429                         ir_node *arg = get_irn_n(irn, i);
 430
 431                         assert(arch_get_irn_reg_class(arg) == co->cls && "Argument not in same register class.");
 432                         if (arg == irn)
 433                                 continue;
 434                         if (nodes_interfere(co->cenv, irn, arg)) {
 435                                 unit->inevitable_costs += co->get_costs(irn, i);
 436                                 continue;
 437                         }
 438
 439                         /* Else insert the argument of the phi to the members of this ou */
 440                         DBG((dbg, LEVEL_1, "\t   Member: %+F\n", arg));
 441
 442                         if (arch_irn_is_ignore(arg))
 443                                 continue;
 444
 445                         /* Check if arg has occurred at a prior position in the arg/list */
 446                         arg_pos = 0;
 447                         for (o=1; o<unit->node_count; ++o) {
 448                                 if (unit->nodes[o] == arg) {
 449                                         arg_pos = o;
 450                                         break;
 451                                 }
 452                         }
 453
 454                         if (!arg_pos) { /* a new argument */
 455                                 /* insert node, set costs */
 456                                 unit->nodes[unit->node_count] = arg;
 457                                 unit->costs[unit->node_count] = co->get_costs(irn, i);
 458                                 unit->node_count++;
 459                         } else { /* arg has occurred before in same phi */
 460                                 /* increase costs for existing arg */
 461                                 unit->costs[arg_pos] += co->get_costs(irn, i);
 462                         }
 463                 }
 464                 unit->nodes = XREALLOC(unit->nodes, ir_node*, unit->node_count);
 465                 unit->costs = XREALLOC(unit->costs, int,      unit->node_count);
 466         } else if (is_Perm_Proj(irn)) {
 467                 /* Proj of a perm with corresponding arg */
 468                 assert(!nodes_interfere(co->cenv, irn, get_Perm_src(irn)));
 469                 unit->nodes = XMALLOCN(ir_node*, 2);
 470                 unit->costs = XMALLOCN(int,      2);
 471                 unit->node_count = 2;
 472                 unit->nodes[0] = irn;
 473                 unit->nodes[1] = get_Perm_src(irn);
 474                 unit->costs[1] = co->get_costs(irn, -1);
 475         } else if (arch_register_req_is(req, should_be_same)) {
 476                 /* Src == Tgt of a 2-addr-code instruction */
 477                 const unsigned other = req->other_same;
 478                 int            count = 0;
 479                 int            i;
 480
 481                 for (i = 0; (1U << i) <= other; ++i) {
 482                         if (other & (1U << i)) {
 483                                 ir_node *o  = get_irn_n(skip_Proj(irn), i);
 484                                 if (arch_irn_is_ignore(o))
 485                                         continue;
 486                                 if (nodes_interfere(co->cenv, irn, o))
 487                                         continue;
 488                                 ++count;
 489                         }
 490                 }
 491
 492                 if (count != 0) {
 493                         int k = 0;
 494                         ++count;
 495                         unit->nodes = XMALLOCN(ir_node*, count);
 496                         unit->costs = XMALLOCN(int,      count);
 497                         unit->node_count = count;
 498                         unit->nodes[k++] = irn;
 499
 500                         for (i = 0; 1U << i <= other; ++i) {
 501                                 if (other & (1U << i)) {
 502                                         ir_node *o  = get_irn_n(skip_Proj(irn), i);
 503                                         if (!arch_irn_is_ignore(o) &&
 504                                                         !nodes_interfere(co->cenv, irn, o)) {
 505                                                 unit->nodes[k] = o;
 506                                                 unit->costs[k] = co->get_costs(irn, -1);
 507                                                 ++k;
 508                                         }
 509                                 }
 510                         }
 511                 }
 512         } else {
 513                 assert(0 && "This is not an optimizable node!");
 514         }
 515
 516         /* Insert the new unit at a position according to its costs */
 517         if (unit->node_count > 1) {
 518                 int i;
 519                 struct list_head *tmp;
 520
 521                 /* Determine the maximum costs this unit can cause: all_nodes_cost */
 522                 for (i=1; i<unit->node_count; ++i) {
 523                         unit->sort_key = MAX(unit->sort_key, unit->costs[i]);
 524                         unit->all_nodes_costs += unit->costs[i];
 525                 }
 526
 527                 /* Determine the minimal costs this unit will cause: min_nodes_costs */
 528                 unit->min_nodes_costs += unit->all_nodes_costs - ou_max_ind_set_costs(unit);
 529                 /* Insert the new ou according to its sort_key */
 530                 tmp = &co->units;
 531                 while (tmp->next != &co->units && list_entry_units(tmp->next)->sort_key > unit->sort_key)
 532                         tmp = tmp->next;
 533                 list_add(&unit->units, tmp);
 534         } else {
 535                 free(unit);
 536         }
 537 }
 538
 539 #ifdef QUICK_AND_DIRTY_HACK
 540
 541 static int compare_ous(const void *k1, const void *k2)
 542 {
 543         const unit_t *u1 = *((const unit_t **) k1);
 544         const unit_t *u2 = *((const unit_t **) k2);
 545         int i, o, u1_has_constr, u2_has_constr;
 546         arch_register_req_t req;
 547
 548         /* Units with constraints come first */
 549         u1_has_constr = 0;
 550         for (i=0; i<u1->node_count; ++i) {
 551                 arch_get_irn_register_req(&req, u1->nodes[i]);
 552                 if (arch_register_req_is(&req, limited)) {
 553                         u1_has_constr = 1;
 554                         break;
 555                 }
 556         }
 557
 558         u2_has_constr = 0;
 559         for (i=0; i<u2->node_count; ++i) {
 560                 arch_get_irn_register_req(&req, u2->nodes[i]);
 561                 if (arch_register_req_is(&req, limited)) {
 562                         u2_has_constr = 1;
 563                         break;
 564                 }
 565         }
 566
 567         if (u1_has_constr != u2_has_constr)
 568                 return u2_has_constr - u1_has_constr;
 569
 570         /* After all, the sort key decides. Greater keys come first. */
 571         return u2->sort_key - u1->sort_key;
 572
 573 }
 574
 575 /**
 576  * Sort the ou's according to constraints and their sort_key
 577  */
 578 static void co_sort_units(copy_opt_t *co)
 579 {
 580         int i, count = 0, costs;
 581         unit_t **ous;
 582
 583         /* get the number of ous, remove them form the list and fill the array */
 584         list_for_each_entry(unit_t, ou, &co->units, units)
 585                 count++;
 586         ous = ALLOCAN(unit_t, count);
 587
 588         costs = co_get_max_copy_costs(co);
 589
 590         i = 0;
 591         list_for_each_entry(unit_t, ou, &co->units, units)
 592                 ous[i++] = ou;
 593
 594         INIT_LIST_HEAD(&co->units);
 595
 596         assert(count == i && list_empty(&co->units));
 597
 598         for (i=0; i<count; ++i)
 599                 ir_printf("%+F\n", ous[i]->nodes[0]);
 600
 601         qsort(ous, count, sizeof(*ous), compare_ous);
 602
 603         ir_printf("\n\n");
 604         for (i=0; i<count; ++i)
 605                 ir_printf("%+F\n", ous[i]->nodes[0]);
 606
 607         /* reinsert into list in correct order */
 608         for (i=0; i<count; ++i)
 609                 list_add_tail(&ous[i]->units, &co->units);
 610
 611         assert(costs == co_get_max_copy_costs(co));
 612 }
 613 #endif
 614
 615 void co_build_ou_structure(copy_opt_t *co)
 616 {
 617         DBG((dbg, LEVEL_1, "\tCollecting optimization units\n"));
 618         INIT_LIST_HEAD(&co->units);
 619         irg_walk_graph(co->irg, co_collect_units, NULL, co);
 620 #ifdef QUICK_AND_DIRTY_HACK
 621         co_sort_units(co);
 622 #endif
 623 }
 624
 625 void co_free_ou_structure(copy_opt_t *co)
 626 {
 627         ASSERT_OU_AVAIL(co);
 628         list_for_each_entry_safe(unit_t, curr, tmp, &co->units, units) {
 629                 xfree(curr->nodes);
 630                 xfree(curr->costs);
 631                 xfree(curr);
 632         }
 633         co->units.next = NULL;
 634 }
 635
 636 /* co_solve_heuristic() is implemented in becopyheur.c */
 637
 638 int co_get_max_copy_costs(const copy_opt_t *co)
 639 {
 640         int i, res = 0;
 641
 642         ASSERT_OU_AVAIL(co);
 643
 644         list_for_each_entry(unit_t, curr, &co->units, units) {
 645                 res += curr->inevitable_costs;
 646                 for (i=1; i<curr->node_count; ++i)
 647                         res += curr->costs[i];
 648         }
 649         return res;
 650 }
 651
 652 int co_get_inevit_copy_costs(const copy_opt_t *co)
 653 {
 654         int res = 0;
 655
 656         ASSERT_OU_AVAIL(co);
 657
 658         list_for_each_entry(unit_t, curr, &co->units, units)
 659                 res += curr->inevitable_costs;
 660         return res;
 661 }
 662
 663 int co_get_copy_costs(const copy_opt_t *co)
 664 {
 665         int i, res = 0;
 666
 667         ASSERT_OU_AVAIL(co);
 668
 669         list_for_each_entry(unit_t, curr, &co->units, units) {
 670                 int root_col = get_irn_col(curr->nodes[0]);
 671                 DBG((dbg, LEVEL_1, "  %3d costs for root %+F color %d\n", curr->inevitable_costs, curr->nodes[0], root_col));
 672                 res += curr->inevitable_costs;
 673                 for (i=1; i<curr->node_count; ++i) {
 674                         int arg_col = get_irn_col(curr->nodes[i]);
 675                         if (root_col != arg_col) {
 676                                 DBG((dbg, LEVEL_1, "  %3d for arg %+F color %d\n", curr->costs[i], curr->nodes[i], arg_col));
 677                                 res += curr->costs[i];
 678                         }
 679                 }
 680         }
 681         return res;
 682 }
 683
 684 int co_get_lower_bound(const copy_opt_t *co)
 685 {
 686         int res = 0;
 687
 688         ASSERT_OU_AVAIL(co);
 689
 690         list_for_each_entry(unit_t, curr, &co->units, units)
 691                 res += curr->inevitable_costs + curr->min_nodes_costs;
 692         return res;
 693 }
 694
 695 void co_complete_stats(const copy_opt_t *co, co_complete_stats_t *stat)
 696 {
 697         bitset_t *seen = bitset_malloc(get_irg_last_idx(co->irg));
 698
 699         memset(stat, 0, sizeof(stat[0]));
 700
 701         /* count affinity edges. */
 702         co_gs_foreach_aff_node(co, an) {
 703                 stat->aff_nodes += 1;
 704                 bitset_set(seen, get_irn_idx(an->irn));
 705                 co_gs_foreach_neighb(an, neigh) {
 706                         if (!bitset_is_set(seen, get_irn_idx(neigh->irn))) {
 707                                 stat->aff_edges += 1;
 708                                 stat->max_costs += neigh->costs;
 709
 710                                 if (get_irn_col(an->irn) != get_irn_col(neigh->irn)) {
 711                                         stat->costs += neigh->costs;
 712                                         stat->unsatisfied_edges += 1;
 713                                 }
 714
 715                                 if (nodes_interfere(co->cenv, an->irn, neigh->irn)) {
 716                                         stat->aff_int += 1;
 717                                         stat->inevit_costs += neigh->costs;
 718                                 }
 719
 720                         }
 721                 }
 722         }
 723
 724         bitset_free(seen);
 725 }
 726
 727 /******************************************************************************
 728    _____                 _        _____ _
 729   / ____|               | |      / ____| |
 730  | |  __ _ __ __ _ _ __ | |__   | (___ | |_ ___  _ __ __ _  __ _  ___
 731  | | |_ | '__/ _` | '_ \| '_ \   \___ \| __/ _ \| '__/ _` |/ _` |/ _ \
 732  | |__| | | | (_| | |_) | | | |  ____) | || (_) | | | (_| | (_| |  __/
 733   \_____|_|  \__,_| .__/|_| |_| |_____/ \__\___/|_|  \__,_|\__, |\___|
 734                   | |                                       __/ |
 735                   |_|                                      |___/
 736  ******************************************************************************/
 737
 738 static int compare_affinity_node_t(const void *k1, const void *k2, size_t size)
 739 {
 740         const affinity_node_t *n1 = (const affinity_node_t*)k1;
 741         const affinity_node_t *n2 = (const affinity_node_t*)k2;
 742         (void) size;
 743
 744         return (n1->irn != n2->irn);
 745 }
 746
 747 static void add_edge(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs)
 748 {
 749         affinity_node_t new_node, *node;
 750         neighb_t        *nbr;
 751         int             allocnew = 1;
 752
 753         new_node.irn        = n1;
 754         new_node.degree     = 0;
 755         new_node.neighbours = NULL;
 756         node = set_insert(affinity_node_t, co->nodes, &new_node, sizeof(new_node), hash_irn(new_node.irn));
 757
 758         for (nbr = node->neighbours; nbr; nbr = nbr->next)
 759                 if (nbr->irn == n2) {
 760                         allocnew = 0;
 761                         break;
 762                 }
 763
 764         /* if we did not find n2 in n1's neighbourhood insert it */
 765         if (allocnew) {
 766                 nbr        = OALLOC(&co->obst, neighb_t);
 767                 nbr->irn   = n2;
 768                 nbr->costs = 0;
 769                 nbr->next  = node->neighbours;
 770
 771                 node->neighbours = nbr;
 772                 node->degree++;
 773         }
 774
 775         /* now nbr points to n1's neighbour-entry of n2 */
 776         nbr->costs += costs;
 777 }
 778
 779 static inline void add_edges(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs)
 780 {
 781         if (! be_ifg_connected(co->cenv->ifg, n1, n2)) {
 782                 add_edge(co, n1, n2, costs);
 783                 add_edge(co, n2, n1, costs);
 784         }
 785 }
 786
 787 static void build_graph_walker(ir_node *irn, void *env)
 788 {
 789         const arch_register_req_t *req;
 790         copy_opt_t                *co  = (copy_opt_t*)env;
 791         int pos, max;
 792
 793         if (get_irn_mode(irn) == mode_T)
 794                 return;
 795         req = arch_get_irn_register_req(irn);
 796         if (req->cls != co->cls || arch_register_req_is(req, ignore))
 797                 return;
 798
 799         if (is_Reg_Phi(irn)) { /* Phis */
 800                 for (pos=0, max=get_irn_arity(irn); pos<max; ++pos) {
 801                         ir_node *arg = get_irn_n(irn, pos);
 802                         add_edges(co, irn, arg, co->get_costs(irn, pos));
 803                 }
 804         } else if (is_Perm_Proj(irn)) { /* Perms */
 805                 ir_node *arg = get_Perm_src(irn);
 806                 add_edges(co, irn, arg, co->get_costs(irn, -1));
 807         } else if (arch_register_req_is(req, should_be_same)) {
 808                 const unsigned other = req->other_same;
 809                 int i;
 810
 811                 for (i = 0; 1U << i <= other; ++i) {
 812                         if (other & (1U << i)) {
 813                                 ir_node *other = get_irn_n(skip_Proj(irn), i);
 814                                 if (!arch_irn_is_ignore(other))
 815                                         add_edges(co, irn, other, co->get_costs(irn, -1));
 816                         }
 817                 }
 818         }
 819 }
 820
 821 void co_build_graph_structure(copy_opt_t *co)
 822 {
 823         obstack_init(&co->obst);
 824         co->nodes = new_set(compare_affinity_node_t, 32);
 825
 826         irg_walk_graph(co->irg, build_graph_walker, NULL, co);
 827 }
 828
 829 void co_free_graph_structure(copy_opt_t *co)
 830 {
 831         ASSERT_GS_AVAIL(co);
 832
 833         del_set(co->nodes);
 834         obstack_free(&co->obst, NULL);
 835         co->nodes = NULL;
 836 }
 837
 838 int co_gs_is_optimizable(copy_opt_t *co, ir_node *irn)
 839 {
 840         affinity_node_t new_node, *n;
 841
 842         ASSERT_GS_AVAIL(co);
 843
 844         new_node.irn = irn;
 845         n = set_find(affinity_node_t, co->nodes, &new_node, sizeof(new_node), hash_irn(new_node.irn));
 846         if (n) {
 847                 return (n->degree > 0);
 848         } else
 849                 return 0;
 850 }
 851
 852 static int co_dump_appel_disjoint_constraints(const copy_opt_t *co, ir_node *a, ir_node *b)
 853 {
 854         ir_node *nodes[]  = { a, b };
 855         bitset_t *constr[] = { NULL, NULL };
 856         int j;
 857
 858         constr[0] = bitset_alloca(co->cls->n_regs);
 859         constr[1] = bitset_alloca(co->cls->n_regs);
 860
 861         for (j = 0; j < 2; ++j) {
 862                 const arch_register_req_t *req = arch_get_irn_register_req(nodes[j]);
 863                 if (arch_register_req_is(req, limited))
 864                         rbitset_copy_to_bitset(req->limited, constr[j]);
 865                 else
 866                         bitset_set_all(constr[j]);
 867
 868         }
 869
 870         return !bitset_intersect(constr[0], constr[1]);
 871 }
 872
 873 /**
 874  * Dump the interference graph according to the Appel/George coalescing contest file format.
 875  * See: http://www.cs.princeton.edu/~appel/coalesce/format.html
 876  * @note Requires graph structure.
 877  * @param co The copy opt object.
 878  * @param f  A file to dump to.
 879  */
 880 static void co_dump_appel_graph(const copy_opt_t *co, FILE *f)
 881 {
 882         be_ifg_t *ifg       = co->cenv->ifg;
 883         int      *color_map = ALLOCAN(int, co->cls->n_regs);
 884         int      *node_map  = XMALLOCN(int, get_irg_last_idx(co->irg) + 1);
 885         ir_graph *irg       = co->irg;
 886         be_irg_t *birg      = be_birg_from_irg(irg);
 887
 888         nodes_iter_t it;
 889         neighbours_iter_t nit;
 890         int n, n_regs;
 891         unsigned i;
 892
 893         n_regs = 0;
 894         for (i = 0; i < co->cls->n_regs; ++i) {
 895                 const arch_register_t *reg = &co->cls->regs[i];
 896                 if (rbitset_is_set(birg->allocatable_regs, reg->global_index)) {
 897                         color_map[i] = n_regs++;
 898                 } else {
 899                         color_map[i] = -1;
 900                 }
 901         }
 902
 903         /*
 904          * n contains the first node number.
 905          * the values below n are the pre-colored register nodes
 906          */
 907
 908         n = n_regs;
 909         be_ifg_foreach_node(ifg, &it, irn) {
 910                 if (arch_irn_is_ignore(irn))
 911                         continue;
 912                 node_map[get_irn_idx(irn)] = n++;
 913         }
 914
 915         fprintf(f, "%d %d\n", n, n_regs);
 916
 917         be_ifg_foreach_node(ifg, &it, irn) {
 918                 arch_register_req_t const *const req = arch_get_irn_register_req(irn);
 919                 if (arch_register_req_is(req, ignore))
 920                         continue;
 921
 922                 int              idx = node_map[get_irn_idx(irn)];
 923                 affinity_node_t *a   = get_affinity_info(co, irn);
 924
 925                 if (arch_register_req_is(req, limited)) {
 926                         for (i = 0; i < co->cls->n_regs; ++i) {
 927                                 if (!rbitset_is_set(req->limited, i) && color_map[i] >= 0)
 928                                         fprintf(f, "%d %d -1\n", color_map[i], idx);
 929                         }
 930                 }
 931
 932                 be_ifg_foreach_neighbour(ifg, &nit, irn, adj) {
 933                         if (!arch_irn_is_ignore(adj) &&
 934                                         !co_dump_appel_disjoint_constraints(co, irn, adj)) {
 935                                 int adj_idx = node_map[get_irn_idx(adj)];
 936                                 if (idx < adj_idx)
 937                                         fprintf(f, "%d %d -1\n", idx, adj_idx);
 938                         }
 939                 }
 940
 941                 if (a) {
 942                         co_gs_foreach_neighb(a, n) {
 943                                 if (!arch_irn_is_ignore(n->irn)) {
 944                                         int n_idx = node_map[get_irn_idx(n->irn)];
 945                                         if (idx < n_idx)
 946                                                 fprintf(f, "%d %d %d\n", idx, n_idx, (int) n->costs);
 947                                 }
 948                         }
 949                 }
 950         }
 951
 952         xfree(node_map);
 953 }
 954
 955 static FILE *my_open(const be_chordal_env_t *env, const char *prefix,
 956                      const char *suffix)
 957 {
 958         FILE *result;
 959         char buf[1024];
 960         size_t i, n;
 961         char *tu_name;
 962         const char *cup_name = be_get_irg_main_env(env->irg)->cup_name;
 963
 964         n = strlen(cup_name);
 965         tu_name = XMALLOCN(char, n + 1);
 966         strcpy(tu_name, cup_name);
 967         for (i = 0; i < n; ++i)
 968                 if (tu_name[i] == '.')
 969                         tu_name[i] = '_';
 970
 971
 972         ir_snprintf(buf, sizeof(buf), "%s%s_%F_%s%s", prefix, tu_name, env->irg, env->cls->name, suffix);
 973         xfree(tu_name);
 974         result = fopen(buf, "wt");
 975         if (result == NULL) {
 976                 panic("Couldn't open '%s' for writing.", buf);
 977         }
 978
 979         return result;
 980 }
 981
 982 void co_driver(be_chordal_env_t *cenv)
 983 {
 984         ir_timer_t          *timer = ir_timer_new();
 985         co_complete_stats_t before, after;
 986         copy_opt_t          *co;
 987         int                 was_optimal = 0;
 988
 989         assert(selected_copyopt);
 990
 991         /* skip copymin if algo is 'none' */
 992         if (selected_copyopt->copyopt == void_algo)
 993                 return;
 994
 995         be_assure_live_chk(cenv->irg);
 996
 997         co = new_copy_opt(cenv, cost_func);
 998         co_build_ou_structure(co);
 999         co_build_graph_structure(co);
1000
1001         co_complete_stats(co, &before);
1002
1003         stat_ev_ull("co_aff_nodes",    before.aff_nodes);
1004         stat_ev_ull("co_aff_edges",    before.aff_edges);
1005         stat_ev_ull("co_max_costs",    before.max_costs);
1006         stat_ev_ull("co_inevit_costs", before.inevit_costs);
1007         stat_ev_ull("co_aff_int",      before.aff_int);
1008
1009         stat_ev_ull("co_init_costs",   before.costs);
1010         stat_ev_ull("co_init_unsat",   before.unsatisfied_edges);
1011
1012         if (dump_flags & DUMP_BEFORE) {
1013                 FILE *f = my_open(cenv, "", "-before.vcg");
1014                 be_dump_ifg_co(f, co, style_flags & CO_IFG_DUMP_LABELS, style_flags & CO_IFG_DUMP_COLORS);
1015                 fclose(f);
1016         }
1017
1018         /* if the algo can improve results, provide an initial solution with heur1 */
1019         if (improve && selected_copyopt->can_improve_existing) {
1020                 co_complete_stats_t stats;
1021
1022                 /* produce a heuristic solution */
1023                 co_solve_heuristic(co);
1024
1025                 /* do the stats and provide the current costs */
1026                 co_complete_stats(co, &stats);
1027                 stat_ev_ull("co_prepare_costs", stats.costs);
1028         }
1029
1030         /* perform actual copy minimization */
1031         ir_timer_reset_and_start(timer);
1032         was_optimal = selected_copyopt->copyopt(co);
1033         ir_timer_stop(timer);
1034
1035         stat_ev_dbl("co_time", ir_timer_elapsed_msec(timer));
1036         stat_ev_ull("co_optimal", was_optimal);
1037         ir_timer_free(timer);
1038
1039         if (dump_flags & DUMP_AFTER) {
1040                 FILE *f = my_open(cenv, "", "-after.vcg");
1041                 be_dump_ifg_co(f, co, style_flags & CO_IFG_DUMP_LABELS, style_flags & CO_IFG_DUMP_COLORS);
1042                 fclose(f);
1043         }
1044
1045         co_complete_stats(co, &after);
1046
1047         if (do_stats) {
1048                 unsigned long long optimizable_costs = after.max_costs - after.inevit_costs;
1049                 unsigned long long evitable          = after.costs     - after.inevit_costs;
1050
1051                 ir_printf("%30F ", cenv->irg);
1052                 printf("%10s %10llu%10llu%10llu", cenv->cls->name, after.max_costs, before.costs, after.inevit_costs);
1053
1054                 if (optimizable_costs > 0)
1055                         printf("%10llu %5.2f\n", after.costs, (evitable * 100.0) / optimizable_costs);
1056                 else
1057                         printf("%10llu %5s\n", after.costs, "-");
1058         }
1059
1060         /* Dump the interference graph in Appel's format. */
1061         if (dump_flags & DUMP_APPEL) {
1062                 FILE *f = my_open(cenv, "", ".apl");
1063                 fprintf(f, "# %llu %llu\n", after.costs, after.unsatisfied_edges);
1064                 co_dump_appel_graph(co, f);
1065                 fclose(f);
1066         }
1067
1068         stat_ev_ull("co_after_costs", after.costs);
1069         stat_ev_ull("co_after_unsat", after.unsatisfied_edges);
1070
1071         co_free_graph_structure(co);
1072         co_free_ou_structure(co);
1073         free_copy_opt(co);
1074 }